Perplexity Teammate orchestrates email, calendars, briefs, and planning

#1 𝕏

Aravind Srinivas said Perplexity Teammate can act as a chief of staff or executive assistant, reporting product-market fit for that role after two days of use. Set up once, it can orchestrate calendar management, email follow-ups, triage, vendor coordination, daily briefs, meeting prep, and work planning instead of handling them in separate Computer sessions.

Also covered by: @Aravind Srinivas

#2 in

Guillermo Rauch shared an approach to generative UI for conversational agents that combines AI SDK, Chat SDK, and json-render to go directly from an LLM to UI on platforms such as Slack and Teams. He believes it will become the dominant Human-Agent Interaction paradigm.

#3 𝕏

Andrej Karpathy created TristanTube from Tristan’s top YouTube channels list, featuring 47K “super nerdy” videos and a TV mode for laid-back watching. Karpathy said he one-shotted it with Claude Code in 3 minutes.

#4 𝕏

Harrison Chase shared LangChain’s blog post “Agents Are Going Async,” which argues that long tasks shouldn’t block the chat.

#5 🐙 blader/humanizer

blader/humanizer - **blader/humanizer** is an MIT-licensed agent skill that rewrites AI-generated prose without changing its meaning, checking 26 writing patterns. The repository includes Claude and Cursor plugin directories, reusable skill files, agent files, and scripts; it has 66,148 stars and a weekly trending gain of 3,067. Worth trying if you need plainer AI-written product copy or documentation.

#6 🐙 anthropics/claude-quickstarts

anthropics/claude-quickstarts - anthropics/claude-quickstarts is an MIT-licensed collection of customizable starter projects for deployable Claude API apps, covering customer support, financial analysis, browser and desktop automation, autonomous coding, and managed agents; a Claude API key is required. It has 19,175 stars and a daily trending gain of 28. Worth trying if you need Claude starter code, though its minimal browser and computer toolset examples are not production code.

#7 𝕏

Yann LeCun commented on differences among image and video self-supervised learning methods, including reconstruction versus joint-embedding approaches. He discussed DINO, JEPA, V-JEPA, anti-collapse methods, and upstream research versus scaling and engineering.

#8 𝕏

Julien Chaumond reshared a NousResearch post introducing Arete-27B, described as Qwen3.8-27B post-trained to reinforce Western values.

#9 in

Zach Davis commented that leaders encountered through work at CXO.dev are concerned about AI spending and proving its effectiveness. He recommended granular spend visibility, session-data-driven optimization, and soft caps as trust-based guardrails, arguing that hard caps discourage experimentation and limit returns.

#10 𝕏

clem 🤗 reshared @DynamicWebPaige’s favorable description of an unspecified benchmark as empirically rigorous, noting that holding the model fixed still produced a ~2x swing in a solve-related metric.

#11 🐙 google-deepmind/mujoco

google-deepmind/mujoco - **google-deepmind/mujoco** is MuJoCo, a general-purpose physics engine shared and maintained by Google DeepMind; the repository includes a C API, Python bindings, browser support, a Unity plug-in, and tutorials. It has 16143 GitHub stars and 26 daily trending gains. Worth trying if you need physics simulation for robotics or machine-learning projects.

#12 ▶️

The one thing holding theoretical CS together just broke... kinda

Fireship